An obstacle detection model

Abstract

The application discloses an obstacle detection model and relates to the technical field of computer vision and deep learning.The application is applied to obstacle perception in an automatic driving system and intelligent logistics and comprises a backbone network, a neck and a prediction head;the backbone network realizes efficient feature representation through linear manifold transformation of 'expansion-feature reconstruction-dimension reduction', and a parallel average pooling enhancement branch is introduced to compensate for inherent feature defects of single maximum pooling and the absence of detailed features;the neck adjusts the sampling position of a convolution kernel through a learned offset, so that the model focuses on the effective feature area of a target, and linear optimization is used to significantly reduce the calculation complexity;an ACmix module and a P2 fine-grained detection layer are deeply fused in the detection head.The application can significantly improve the decision safety and operation efficiency of intelligent equipment in a complex dynamic environment.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning technology, and more specifically to an obstacle detection model. Background Technology

[0002] Current obstacle detection technology has shifted entirely from traditional image processing methods to deep learning-based target detection frameworks. These primarily include two-stage detection algorithms, such as the R-CNN series, which achieve high accuracy by generating candidate regions, and single-stage detection algorithms, such as the YOLO and SSD series, which treat detection as a regression problem to improve processing speed. In the fields of autonomous driving (ADS) and smart logistics (SL), to address the complex environmental perception requirements, existing technologies also encompass point cloud clustering and filtering methods based on 2D LiDAR, multimodal data fusion recognition algorithms combining visible light and infrared images, and perception frameworks integrating deep convolutional neural networks (CNNs) and position probability functions. These methods, by utilizing multi-source sensor information and deep feature extraction techniques, aim to provide obstacle recognition support for path planning and behavioral decision-making for automated tractors, flatbed trucks, and autonomous mobile robots (AMRs) in urban roads and industrial warehousing environments.

[0003] While existing technologies have made some progress in conventional scenarios, they still face significant challenges in complex mixed environments where autonomous vehicles, logistics robots, and pedestrians coexist. First, traditional single-stage algorithms often sacrifice the ability to detect small objects while ensuring real-time performance, making it easy to miss obstacles such as scattered debris with sparse pixels in factory aisles. Second, in scenarios with dense crowds or stacked shelves, severe occlusion and bounding box overlap lead to the loss of key visual features, making it difficult for existing detectors to reconstruct the complete target shape. Furthermore, traditional models have weak perception of geometric deformations, and in low-light conditions such as dimly lit warehouses, blurred edges and color distortions in images are amplified layer by layer by nonlinear activation functions, contaminating deep semantic features and causing a significant drop in detection accuracy. In summary, existing technologies lack a unified perception and recognition framework that can simultaneously achieve high robustness and real-time accuracy in extreme environments, with small targets, and in occluded scenarios. Summary of the Invention

[0004] In view of this, the present invention provides an obstacle detection model to solve the problems existing in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An obstacle detection model is applied to obstacle perception in autonomous driving systems and intelligent logistics, including a backbone network, neck, and predictive head. The backbone network achieves efficient feature representation through a linear manifold transformation of "expansion-feature reconstruction-dimensionality reduction" and introduces a parallel average pooling enhancement branch to compensate for the inherent feature defects and lack of detailed features of a single max pooling. The neck adjusts the sampling position of the convolution kernel through the learned offset, so that the model focuses on the effective feature region of the target, while significantly reducing the computational complexity through linear optimization. The detection head deeply integrates the ACmix module with the P2 fine-grained detection layer.

[0006] Optionally, the backbone network employs an efficient feature extraction module to achieve efficient feature representation through a linear manifold transformation of "expansion-feature reconstruction-dimensionality reduction"; wherein, the efficient feature extraction module includes: Using convolution to expand features in high dimension transitions the data into an expanded stage of high-dimensional representation; A depthwise convolutional stage is used to perform spatial feature extraction in a high-dimensional space using depthwise separable convolutions. A channel feature recalibration process that combines attention mechanisms to model and recalibrate the feature dependencies between channels; In addition, to enhance the model's generalization ability, Dropout and linear bottleneck operations are added before the residual connections.

[0007] Optional, expansion stage, processed feature map The number of channels is extended to the original input. Six times that, the formula is as follows: ; in, express Convolution operation, Representatives criticized normalization. This represents the Swish activation function.

[0008] Optionally, the expression for the depthwise convolution stage is as follows: ; in, Indicates the number of output channels. This represents the output of this stage. This represents a depthwise separable convolution operation.

[0009] Optionally, the attention mechanism processes data in two phases: a compression phase and an activation phase. During the compression phase, the module uses global average pooling to compress the global spatial distribution of each feature channel into a single descriptor. : ; in It is the first Global statistics for each channel It is the first feature map output by the depthwise convolution. One channel, Indicates spatial dimension; The weight distribution is generated during the incentive phase using the following formula. : ; in and It is the weight matrix of the fully connected layer designed for channel-dimensional modulation; Represents the ReLU activation function. This represents the Sigmoid activation function. Represents a global statistical vector. It is a weight vector; Performing scaling operations will calculate the weights Apply back to the original feature map: ; in It is a recalibrated feature map. This represents the scalar weight of the channel.

[0010] Optional, the final output of Dropout and linear bottleneck operations Formulated as: ; in," "Indicates residual connection, It is a random variable that follows a Bernoulli distribution, representing the Dropout mask. This represents the weight matrix of the reconvolution; " indicates a convolution operation.

[0011] Optionally, a parallel average pooling enhancement branch is introduced through a multi-dimensional feature context aggregation network to compensate for the inherent feature deficiencies and lack of detailed features of single max pooling. The process formula of the multi-dimensional feature context aggregation network is as follows: ; in Indicates the first branch in the two branches The output of each pooling layer; Globally stitch together the multi-scale features obtained from the dual paths, using... Convolutional layers perform dimensionality reduction and information exchange across channels, as shown below: ; Indicates the input feature map, This represents the SiLU activation function.

[0012] Optionally, the neck area employs an adaptive feature enhancement module, expressed as follows: ; in This represents a traditional sampling grid. This represents the initial mesh of LDConv; Learnable offset tensors are generated through convolution. Add it to the original coordinates In this way, the modified sampling coordinates can be derived. As shown below: ; in These are the current position coordinates; Bilinear interpolation is used during the resampling process, as follows: ; in Represents the four nearest pixels, It's a pixel value. These are coordinate components; Feature extraction is achieved through reshaping transformation and linear sampling, generating an output feature map: ; in, Represents the convolution parameters. That is the corresponding pixel value.

[0013] Optionally, the detection head adopts a high-granularity feature-aware detection head, including a feature projection stage and a dual-path aggregation stage; In the projection stage, the input features First, after three independent Convolutional layer processing, mapping to containing intermediate feature set of each component The computational process is decoupled into two highly complementary paths: the self-attention path aims to... Divided into We use functional groups to extract global context dependencies, and the formal expression is as follows: ; in Indicates the first Learnable projection matrix for each attention head Defined in pixels Centered A local neighborhood, symbol Indicates from The connection operation of the output of each independent attention head This indicates the calculation of attention weights; The product path captures fine-grained spatial geometric features through local operators and processes them using a weight kernel generated by a set of fully connected layers. And perform feature shifting and summation as follows: ; in Indicates the convolution kernel weights. Represents the discrete sampling positions within the convolution kernel; function Perform a spatial translation operation to align the feature tensors with their respective grid coordinates; The feature representations of the two paths are respectively processed through learnable coefficients. and Weighted and scaled values ​​are then integrated via element-wise fusion before the P2 prediction layer to generate the final high-resolution output. .

[0014] As can be seen from the above technical solutions, compared with the prior art, the present invention provides an obstacle detection model that can not only accurately identify vehicles, pedestrians and traffic signs in urban road environments, but also be deployed in autonomous mobile robots (AMRs) in factory or warehouse environments to detect stacks, scattered debris and moving people; especially for extreme visual scenarios such as dense crowds, small objects, severe occlusion and low light, the present invention provides a highly reliable target detection solution that can significantly improve the decision-making safety and operational efficiency of intelligent equipment in complex dynamic environments. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the obstacle detection model structure provided by the present invention; Figure 2 Here is a diagram of the MBConv structure; Figure 3 A comparison chart of different pyramid pooling methods; Figure 4 A flowchart for linear deformable convolution; Figure 5 A diagram showing the structure of a precise small target detector; Figure 6A visual comparison chart of ODDet and YOLO series; Figure 7 A comparative visualization of the heatmap for the OD dataset; Figure 8 This is a visualization comparison of experiments using the OPS dataset. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This invention discloses an obstacle detection model applied to obstacle perception in autonomous driving systems and intelligent logistics, such as... Figures 1-8 As shown, it includes the backbone network, neck, and prediction head; The backbone network achieves efficient feature representation through a linear manifold transformation of "expansion-feature reconstruction-dimensionality reduction" and introduces a parallel average pooling enhancement branch to compensate for the inherent feature defects and lack of detailed features of a single max pooling. The neck adjusts the sampling position of the convolution kernel through the learned offset, so that the model focuses on the effective feature region of the target, while significantly reducing the computational complexity through linear optimization. The detection head deeply integrates the ACmix module with the P2 fine-grained detection layer.

[0019] Consistent with the modular design paradigm of the classic YOLO framework, the obstacle detection model mainly consists of the input image and includes three components: the backbone network, the neck, and the prediction head. Its structural details are as follows... Figure 1 As shown, the basic backbone network employs EFDF, which integrates MBConv and DPSPPF to achieve lightweight and efficient feature extraction while compensating for the loss of detailed features. In the neck stage, LDConv is introduced to replace standard convolutions, significantly enhancing the framework's ability to extract features from irregularly shaped and partially occluded targets. Finally, an Ashead branch is integrated into the detection head to enhance the perception of subtle textures, thereby ensuring superior detection performance.

[0020] To address the complexity of dynamic urban traffic and highly variable industrial logistics scenarios, this paper introduces MBConv as the main feature extraction module of the backbone network. It achieves efficient feature representation through a linear manifold transformation involving "expansion-feature reconstruction-dimensionality reduction." Its architecture is as follows: Figure 2 As shown, the mathematical processing flow of MBConv mainly includes four key stages: At this stage, MBConv uses Convolution expands the feature channels in a high dimension, transitioning the data to a high-dimensional representation. The processed feature map... The number of channels is extended to the original input. Six times, as shown in the formula: (1) in, express Convolution operation, Representatives criticized normalization. This represents the Swish activation function.

[0021] To expand the model's receptive field to capture multi-scale obstacle features while maintaining low computational complexity, MBConv uses depthwise separable convolutions to perform spatial feature extraction in high-dimensional space. The computational cost of this operation is only a fraction of that of standard convolutions. ,in This indicates the number of output channels. The process at this stage can be described as follows: (2) in, This represents the output of this stage. This represents a depthwise separable convolution operation.

[0022] MBConv incorporates a Squeeze-and-Excitation (SE) attention mechanism to model and recalibrate feature dependencies between channels. The SE module divides the processing into two distinct phases: compression and excitation. In the compression phase, the module employs global average pooling to compress the global spatial distribution of each feature channel into a single descriptor. : (3) in It is the first Global statistics for each channel It is the first feature map output by the depthwise convolution. One channel, Indicates spatial dimension.

[0023] Subsequently, the weight distribution is generated during the incentive phase using the following formula. : (4) in and It is the weight matrix of the fully connected layer designed for channel-dimensional modulation; Represents the ReLU activation function. This represents the Sigmoid activation function. Represents a global statistical vector. It is a weight vector.

[0024] Finally, performing the scaling operation will calculate the weights. Apply back to the original feature map: (5) in It is a recalibrated feature map. This represents the scalar weight of the channel.

[0025] To enhance the model's generalization ability, a Dropout operation was added to the MBConv module before the residual connections. Final output. Formulated as: (6) in," "Indicates residual connection, It is a random variable that follows a Bernoulli distribution, representing the Dropout mask. This represents the weight matrix of the deconvolution. " indicates a convolution operation.

[0026] In summary, integrating MBConv into the OCDet framework effectively addresses the core challenges of autonomous obstacle detection by balancing high-dimensional feature preservation and computational efficiency. MBConv leverages its squeeze-excitation recalibration and inverse residual flow to effectively mitigate the sensitivity degradation issue associated with small and occluded objects in complex driving environments. Furthermore, the introduction of stochastic depth ensures robust generalization capabilities across various road and industrial logistics scenarios, making it an ideal computational unit for real-time, safety-critical perception tasks.

[0027] In crowded environments, obstacles often exhibit incomplete feature representations due to mutual occlusion or uneven lighting. Traditional SPP and SPPF only employ max pooling, which, while extracting salient target features, often loses subtle background details and spatial consistency. To address this, this embodiment designs a DPSPPF module, which introduces a parallel average pooling enhancement branch to compensate for the inherent feature deficiencies of single max pooling, thereby improving the model's ability to reconstruct and recognize occluded targets.

[0028] like Figure 3As shown in the structural comparison, DPSPPF achieves a better balance between the depth and breadth of multi-scale feature aggregation compared to previous algorithms. SPP captures the receptive field by parallel pooling with different kernel sizes, but at the cost of lower computational efficiency; SPPF optimizes inference speed by cascading small kernel pooling, but its single-path max pooling mechanism is prone to losing feature information in complex contexts. DPSPPF inherits the cascading efficiency of SPPF and further introduces a parallel average pooling path. This dual-path design not only preserves the significant edge features of obstacles, but also captures smoother global contextual information through average pooling operations.

[0029] The process is shown in Formula 7: (7) in Indicates the first branch in the two branches The output of each pooling layer.

[0030] Subsequently, the module globally stitches together the multi-scale features obtained from the dual paths, and finally uses... Convolutional layers perform dimensionality reduction and information exchange across channels, as shown below: (8) here, Indicates the input feature map, This represents the SiLU activation function.

[0031] In autonomous driving tasks, compared with SPP and SPPF, DPSPPF significantly improves the detection accuracy of occluded targets and alleviates the information collapse problem inherent in the single pooling mechanism during feature downsampling.

[0032] By integrating MBConv with the DPSPPF module, this embodiment develops an EFDF backbone network designed to address the conflict between limited computational resources and feature degradation in complex environments such as ADS and SL. The advantage of EFDF lies in its comprehensive feature reconstruction and aggregation process: First, it utilizes the inverse residual flow of MBConv for efficient high-dimensional extraction, ensuring the linear integrity of semantic information within the high-dimensional manifold. Subsequently, the DPSPPF module is embedded into deeper layers, dynamically capturing multi-scale global context through its parallel max-pooling and average-pooling architecture. This synergy effectively mitigates the sensitivity degradation to small-scale and occluded targets during downsampling and significantly improves discrimination accuracy and generalization robustness under heterogeneous conditions such as dynamic urban traffic and unpredictable industrial logistics environments through cascaded optimization.

[0033] Traditional convolution operators sample on a fixed grid, which limits the model's ability to represent objects with complex geometric deformations. Therefore, this embodiment introduces LDConv, which adjusts the sampling position of the convolution kernel using learned offsets, enabling the model to focus on the effective feature regions of the target while significantly reducing computational complexity through linear optimization.

[0034] The implementation process of LDConv is as follows: Figure 4 As shown. With a Taking the convolution kernel as an example, its initial sampling grid is similar to that of traditional convolution; however, due to the irregular sampling configuration, the center point... It is uniformly defined as the top left corner, as shown in Formula 9.

[0035] (9)

[0036] in This represents a traditional sampling grid. This represents the initial mesh of LDConv.

[0037] Subsequently, LDConv generates learnable offset tensors through convolution. Add it to the original coordinates In this way, the modified sampling coordinates can be derived. As shown in 10: (10) in These are the current position coordinates.

[0038] To address the instability of offset coordinates, bilinear interpolation is used during the resampling process, as follows: (11) in Represents the four nearest pixels, It's a pixel value. These are coordinate components.

[0039] Finally, feature extraction is achieved through reshaping transformation and linear sampling to generate the output feature map: (12) here, Represents the convolution parameters. That is the corresponding pixel value.

[0040] In integrated ADS and SL environments, obstacles such as pedestrians, cyclists, and turning vehicles have diverse shapes. LDConv utilizes parameterized, learnable offsets to enable convolutional kernels to fit the actual contours, significantly enhancing feature extraction of non-rigid objects. Even in densely populated areas or when distant images are blurry and feature points are sparse, LDConv can concentrate sampling points on key areas such as wheels, headlights, or torsos, effectively suppressing background noise and significantly improving feature recognition sensitivity under complex lighting conditions.

[0041] In object detection architectures, traditional detection heads typically use a uniform convolutional structure to process feature maps, ignoring the inherent differences between different feature levels (high-level semantics vs. low-level details). To address this, this embodiment proposes Ashead, a task decoupling component specifically designed for complex scenarios. Its core innovation lies in the deep fusion of the ACmix module and the P2 fine-grained detection layer.

[0042] like Figure 5 As shown, the Ashead framework divides the operation process into two distinct stages: feature projection and two-path aggregation. In the projection stage, the input features... First, after three independent Convolutional layer processing, mapping to containing intermediate feature set of each component This shared projection layer provides a lightweight foundation for subsequent parallel branch computations. The computational flow is then decoupled into two highly complementary paths: a self-attention path designed to... Divided into The system uses three functional groups to extract global context dependencies. These functional groups represent queries, keys, and values, and their formal expressions are as follows: (13) in Indicates the first Learnable projection matrix for each attention head Defined in pixels Centered A local neighborhood, symbol Indicates from The connection operation of the output of each independent attention head This represents the calculation of attention weights.

[0043] Meanwhile, the convolutional path captures fine-grained spatial geometric features through local operators. This path utilizes weight kernels generated by a set of fully connected layers for processing. And perform feature shifting and summation, as shown in Equation 14: (14) in Indicates the convolution kernel weights. Represents the discrete sampling positions within the convolution kernel; function Perform a spatial translation operation to align the feature tensors with their respective grid coordinates.

[0044] Finally, the feature representations of the two paths are processed using learnable coefficients. and Weighted and scaled values ​​are then integrated via element-wise fusion before the P2 prediction layer to generate the final high-resolution output. (15) ASHead's structural design endows the detection head with exceptional flexibility. Especially under extreme conditions such as low signal-to-noise ratio (e.g., low-light environments) or sparse features (e.g., distant targets), ASHead can autonomously suppress background noise and filter out the most discriminative spatial and channel information. This combination of "adaptive operators and high-resolution branches" not only effectively alleviates the problem of missed detections caused by feature degradation but also significantly improves ODDet's classification accuracy and localization sensitivity in complex urban road and industrial logistics scenarios.

[0045] The effectiveness of this invention is further illustrated by the following specific experiment: To verify the effectiveness of ODDet in obstacle detection tasks, this embodiment uses the publicly available ODD dataset on GitHub. This dataset was compiled by Wu et al. using multiple datasets, including VOC, COCO, and TT100K, and supplemented with images collected in the field by the authors. The dataset contains 7915 images, which in this embodiment are divided into 5066 training images, 1266 validation images, and 1583 test images. It includes 15 different categories of urban obstacles: stop signs, pedestrians, bicycles, buses, trucks, cars, motorcycles, reflective cones, trash cans, warning posts, spherical barriers, utility poles, dogs, tricycles, and fire hydrants.

[0046] Furthermore, to evaluate the model's robustness in different public environments, this embodiment extends the experiments to the publicly available OPS dataset on Kaggle. The OPS dataset contains 22 common obstacles in public areas, ranging from directional signs to complex road entities such as puddles and street vendors. To verify the effectiveness of obstacle detection, this embodiment divides the dataset into a training set and a test set in an 8:2 ratio, containing 2684 and 671 images respectively. All experimental results in this paper are evaluated based on the test set.

[0047] In this embodiment, all experimental steps were performed on a workstation running Ubuntu 20.04 with an Intel Xeon E5-2683 v4 CPU. An NVIDIA GeForce RTX 2080 Ti GPU was used to accelerate training. The software stack was built using CUDA 12.6, the PyTorch 2.4.1 deep learning framework, and Python 3.10.16. Key training parameters were set as follows: initial learning rate of 0.01, batch size of 16, and input image resolution of 640×640. Other training hyperparameters used the default values ​​from the Ultralytics YOLOv8 official codebase.

[0048] To objectively evaluate the detection performance and model complexity of the ODDet model, the experiment used precision (P), recall (R), mean average precision (mAP), number of floating-point operations (FLOPs), and number of parameters (Params) as key evaluation metrics. Their calculation formulas are defined as follows: (16) Where TP represents the number of true positives, FP represents the number of false positives, FN represents the number of false negatives, P represents precision, and R represents recall. This represents the average precision rate, i.e., the area under the PR curve. This represents the average precision across all categories.

[0049] To comprehensively evaluate the detection performance of ODDet, this embodiment benchmarked it against the YOLO series models, and the results are summarized in Table 1.

[0050] Table 1: Performance Comparison of ODEt and YOLO Models

[0051] Experimental results demonstrate that ODEt exhibits superior performance, achieving mAP@50 and mAP@50:95 scores of 0.710 and 0.509, respectively, significantly outperforming other comparative models. Specifically, compared to the baseline model YOLOv8n, ODEt achieves significant improvements of 5.0% and 5.2% in mAP@50 and mAP@50:95 scores, respectively. Despite a slight increase in model size, comparative analysis with YOLOv8s shows that ODEt achieves a better balance between detection accuracy and model complexity: while improving accuracy (2.1% improvement in mAP@50 and mAP@50:95), the number of parameters is reduced by 33.3%, and the number of floating-point operations (FLOPs) is reduced by 63.9%. This ability to achieve superior accuracy with a relatively compact model size highlights ODEt's capabilities in high-precision multi-class obstacle detection.

[0052] Table 2: Performance Benchmark Comparison of ODDet with Various Mainstream Architectures

[0053] Table 2 presents a comparative analysis of ODEt with several mainstream detection models. Experimental results show that ODEt significantly outperforms Faster R-CNN in mAP@50:95. Crucially, ODEt demonstrates a significant advantage in lightweight design: its parameter count is only 17.9% of Faster R-CNN's, and its computational complexity (FLOPs) is only 5.7%. Although Cascade R-CNN's detection accuracy is comparable to ODEt's, its large model size and high computational cost severely hinder its practical deployment. Furthermore, even the state-of-the-art DEIM model cannot match the performance of the ODEt framework. In summary, ODEt achieves low computational resource consumption and low model complexity while maintaining high detection accuracy, demonstrating excellent deployment capabilities and strong practical performance.

[0054] To verify the effectiveness of EFDF, this embodiment was compared with various lightweight backbone networks, and the results are summarized in Table 3.

[0055] Table 3: Comparative experiments of mainstream networks on the OD dataset to verify the efficiency of EFDF.

[0056] Experimental results show that while ShuffleNetV2 has advantages in reducing parameters and FLOPs, its detection accuracy is significantly reduced, failing to meet the requirements of high-precision obstacle detection. MobileNetV2 significantly improves detection accuracy, achieving 0.678 mAP@50 and 0.482 mAP@50:95. However, the surge in model size and computational load hinders its efficient deployment. Furthermore, this embodiment also evaluates mainstream Vision Transformer (ViT) architectures, such as MobileViT and RepViT. The results show that although these variants improve accuracy, their increased number of parameters and computational overhead do not reflect the model's efficiency. In contrast, EFDF achieves significant superiority across all four core detection metrics, while drastically reducing FLOPs by 32.1% to only 5.5 G. Despite a slight increase in the number of parameters, its superior detection performance fully reflects the efficiency of its feature extraction, successfully achieving an effective balance between detection performance and model size.

[0057] Table 4: Comparative experiments of different types of detection heads on the OD dataset

[0058] Table 4 evaluates the impact of four different detector head combinations on model performance, where P345 represents the configuration using detector heads P3, P4, and P5. Experimental results show that introducing the P2 detector head effectively improves detection performance, increasing mAP@50 and mAP@50:95 by 1.0% and 1.3%, respectively, with a FLOPs increase of 4.1 G. Conversely, adding the P6 detector head did not bring significant performance improvement but instead introduced 2.4 million redundant parameters. Furthermore, removing the P5 detector head leads to a significant decrease in mAP@50:95 performance. After comprehensive evaluation, this embodiment selects the P2345 combination as the optimal detector head configuration for ODDet, achieving the best balance between multi-scale coverage and overall performance.

[0059] This embodiment integrates the ACmix module into Ashead. To verify its effectiveness, a series of comparative experiments were conducted, replacing ACmix with various mainstream modules, including attention mechanism modules and plug-and-play convolutional blocks. The results are summarized in Table 5.

[0060] Table 5: Comparative experiments of replacing different modules in ASHead on the OD dataset

[0061] Specifically, compared to Ashead, the Global Attention (GAM) mechanism resulted in a 0.3% decrease in both mAP@50 and mAP@50:95, while increasing computational cost by 0.9 GFLOPs. Although the computational complexity of the Inverse Residual Moving Block (IRMB) was reduced, it still lagged behind Ashead by 0.3% in the mAP@50:95 metric. Furthermore, while Multi-Level Context Aggregation (MLCA) and Multi-Source Domain Adaptation (MSDA) focused on multi-scale feature fusion and multi-label generalization, they did not achieve significant accuracy improvements over Ashead. In conclusion, the organic combination of ACmix and the P2 head enabled Ashead to achieve superior performance and efficiency in obstacle detection tasks.

[0062] Table 6: Comparative experiments of different mainstream modules in the backbone and neck networks on the OD dataset.

[0063] Table 6 presents the comparative experimental results of applying the DPSPPF and LDConv modules to the backbone and neck regions, respectively, against other mainstream modules, aiming to demonstrate the effective contribution of these two modules to model performance. In the backbone region, the DPSPPF module significantly improves detection accuracy, outperforming standard SPPF with only a slight increase in parameters and FLOPs. Furthermore, it significantly outperforms all other mainstream methods, including SimSPPF, SPPCSPC, and SPPELAN, while the overall model complexity difference is negligible. In the neck region, LDConv achieves an excellent balance between performance and efficiency. Compared to lightweight convolutional modules such as GSConv and DWConv, LDConv leverages its adaptive convolutional characteristics to effectively improve detection accuracy with only a minimal number of additional parameters. Moreover, although the ODConv and SAConv modules are also compared in the same dimension, their parameter counts are higher than LDConv, and their performance improvement in detection accuracy is negligible. In summary, these comparative experiments strongly demonstrate the superiority of DPSPPF in enhancing backbone network feature representation, and the effectiveness of LDConv in achieving efficient and accurate feature extraction at the network neckline.

[0064] To determine the optimal configuration of the LDConv module, this embodiment investigated the effect of changing the number of convolutional kernels K on detection accuracy; the results are shown in Table 7.

[0065] Table 7: Performance comparison of LDConv on the OD dataset with different kernel sizes (K)

[0066] Experiments show that the model performance peaks when the number of convolutional kernels is set to K=9, achieving the highest overall results: mAP@50 reaches 0.710, and mAP@50:95 reaches 0.509. While theoretically increasing the K value might further improve accuracy, subsequent experiments with K=10 and K=11 clearly demonstrate that LDConv's performance saturates and begins to decline at K=9, while the number of parameters continues to increase. Considering the trade-off between detection accuracy, parameter efficiency, and model complexity, this embodiment ultimately selects the LDConv module with K=9 as the final model configuration.

[0067] To verify the generalization performance of the ODDet model in obstacle detection tasks, this embodiment conducted cross-dataset generalization experiments using the OPS dataset. As shown in Table 8, this embodiment selected eight main obstacle categories for detailed detection performance analysis.

[0068] Table 8: Comparison of domain generalization experiments of ODDet on the OPS dataset

[0069] Experimental results show that ODDet achieves best detection performance in key categories such as puddles, chairs, fences, and trees. Based on a comprehensive analysis of all 22 categories, ODDet demonstrates the best overall detection performance on the OPS dataset, achieving an mAP@50 of 0.855 and an mAP@50:95 of 0.617. Compared to the baseline model YOLOv8n, ODDet significantly improves performance by 2.4% on mAP@50 and by 3.2% on mAP@50:95. Furthermore, compared to the more powerful YOLOv8s, ODDet still achieves performance improvements of 1% and 1.6%, respectively. This fully demonstrates ODDet's superior generalization ability in obstacle detection tasks when facing data from unknown domains.

[0070] To visually demonstrate the superior performance of the ODDet model in obstacle detection tasks, this embodiment presents three sets of visualization experiments, as shown below. Figure 6 , Figure 7 and Figure 8 As shown.

[0071] Figure 6 This visualization compares the qualitative detection performance of ODDet and YOLO series models on the OD dataset. It is based on five randomly selected images containing dense, small, occluded, and low-light objects. The results show that ODDet can detect smaller targets while significantly reducing false negatives and false positives. Furthermore, ODDet exhibits higher confidence and better detection results in occluded and low-light environments.

[0072] Figure 7 A heatmap comparison of ODDet and the baseline model YOLOv8n is presented. This heatmap is generated based on Gradient Weighted Class Activation Mapping (Grad-CAM) [selvaraju2017grad] and uses four random images. The results show that YOLOv8n exhibits highly scattered attention in obstacle detection, frequently misfocusing on background or irrelevant objects, leading to low detection efficiency. In contrast, ODDet demonstrates superior focusing ability and robustness, accurately targeting smaller or real obstacle regions for accurate detection.

[0073] also, Figure 8 Visual comparisons of ODDet and YOLOv8n on different datasets are provided. The results strongly indicate that YOLOv8n has high false negative and false positive rates in different domains, especially when detecting microscopic and occluded obstacles.

[0074] In summary, all visualization results consistently demonstrate that ODDet possesses superior detection performance, powerful decision-making capabilities, and excellent cross-dataset generalization ability, effectively addressing the key challenges inherent in obstacle detection tasks for autonomous driving.

[0075] Table 9: Ablation experiments of the proposed module on the OCDet model and the OD dataset.

[0076] Table 9 provides a quantitative evaluation of the contribution of each module to the ODEt model. First, replacing the baseline backbone network with the EFDF module improved the model's accuracy to 0.789, with significant improvements of 3.0% and 2.9% in mAP@50 and mAP@50:95, respectively. This strongly demonstrates that EFDF effectively enhances feature extraction and multi-scale detail enhancement. Subsequently, to improve small object detection performance, this embodiment integrates the ASHead module in the detection stage. Experiment 6 shows that the combination of EFDF and ASHead significantly improves recall; by fusing local and global information, it greatly enhances the model's stability and generalization ability. Finally, in Experiment 8, this embodiment integrates the LDConv module into the neck network. Utilizing its dynamic adaptive mechanism, this module enables the model to better adapt to different dataset features, thereby enhancing the model's robustness across different categories. Ultimately, ODEt achieved an mAP@50 of 0.710 and an mAP@50:95 of 0.509, representing improvements of 5.0% and 5.2% compared to the baseline model, respectively. This result achieves an excellent balance between detection accuracy and model size, demonstrating the superior performance and synergistic integration of the proposed module.

[0077] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0078] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An obstacle detection model, characterized in that, Obstacle perception applied in autonomous driving systems and smart logistics, including backbone networks, neck and predictive heads; The backbone network achieves efficient feature representation through a linear manifold transformation of "expansion-feature reconstruction-dimensionality reduction" and introduces a parallel average pooling enhancement branch to compensate for the inherent feature defects and lack of detailed features of a single max pooling. The neck adjusts the sampling position of the convolution kernel through the learned offset, so that the model focuses on the effective feature region of the target, while significantly reducing the computational complexity through linear optimization. The detection head deeply integrates the ACmix module with the P2 fine-grained detection layer.

2. The obstacle detection model according to claim 1, characterized in that, The backbone network employs an efficient feature extraction module to achieve efficient feature representation through a linear manifold transformation of "expansion-feature reconstruction-dimensionality reduction"; the efficient feature extraction module includes: Using convolution to expand features in high dimension transitions the data into an expanded stage of high-dimensional representation; A depthwise convolutional stage is used to perform spatial feature extraction in a high-dimensional space using depthwise separable convolutions. A channel feature recalibration process that combines attention mechanisms to model and recalibrate the feature dependencies between channels; In addition, to enhance the model's generalization ability, Dropout and linear bottleneck operations are added before the residual connections.

3. The obstacle detection model according to claim 2, characterized in that, During the expansion phase, the processed feature map The number of channels is extended to the original input. Six times, the formula is as follows: ； in, express Convolution operation, Representatives criticized normalization. This represents the Swish activation function.

4. The obstacle detection model according to claim 2, characterized in that, The expression for the process of the depthwise convolution stage is as follows: ； in, Indicates the number of output channels. This represents the output of this stage. This represents a depthwise separable convolution operation.

5. The obstacle detection model according to claim 1, characterized in that, The attention mechanism processes data in two phases: compression and activation. During the compression phase, the module uses global average pooling to compress the global spatial distribution of each feature channel into a single descriptor. : ； in It is the first Global statistics for each channel It is the first feature map output by the depthwise convolution. One channel, Indicates spatial dimension; The weight distribution is generated during the incentive phase using the following formula. : ； in and It is the weight matrix of the fully connected layer designed for channel-dimensional modulation; Represents the ReLU activation function. This represents the Sigmoid activation function. Represents a global statistical vector. It is a weight vector; Performing scaling operations will calculate the weights Apply back to the original feature map: ； in It is a recalibrated feature map. This represents the scalar weight of the channel.

6. The obstacle detection model according to claim 1, characterized in that, The final output of Dropout and linear bottleneck operations Formulated as: ； in," "Indicates residual connection, It is a random variable that follows a Bernoulli distribution, representing the Dropout mask. This represents the weight matrix of the reconvolution; " indicates convolution operation.

7. The obstacle detection model according to claim 1, characterized in that, A parallel average pooling enhancement branch is introduced through a multi-dimensional feature context aggregation network to compensate for the inherent feature deficiencies and lack of detailed features of single max pooling. The process formula of the multi-dimensional feature context aggregation network is as follows: ； in Indicates the first of the two branches The output of each pooling layer; Globally stitch together the multi-scale features obtained from the dual paths, using... Convolutional layers perform dimensionality reduction and information exchange across channels, as shown below: ； Indicates the input feature map, This represents the SiLU activation function.

8. The obstacle detection model according to claim 1, characterized in that, The neck region employs an adaptive feature enhancement module, expressed as follows: ； in This represents a traditional sampling grid. This represents the initial mesh of LDConv; Learnable offset tensors are generated through convolution. Add it to the original coordinates In this way, the modified sampling coordinates can be derived. As shown below: ； in These are the current position coordinates; Bilinear interpolation is used during the resampling process, as follows: ； in Represents the four nearest pixels, It's a pixel value. These are coordinate components; Feature extraction is achieved through reshaping transformation and linear sampling, generating an output feature map: ； in, Represents the convolution parameters. That is the corresponding pixel value.

9. The obstacle detection model according to claim 1, characterized in that, The detection head uses a high-granularity feature sensing detection head. This includes the feature projection stage and the dual-path aggregation stage; In the projection stage, the input features First, after three independent Convolutional layer processing, mapping to containing intermediate feature set of each component The computational process is decoupled into two highly complementary paths: the self-attention path aims to... Divided into We use functional groups to extract global context dependencies, and the formal expression is as follows: ； in Indicates the first Learnable projection matrix for each attention head Defined in pixels Centered A local neighborhood, symbol Indicates from The connection operation of the output of each independent attention head This indicates the calculation of attention weights; The product path captures fine-grained spatial geometric features through local operators and processes them using a weight kernel generated by a set of fully connected layers. And perform feature shifting and summation as follows: ； in Indicates the convolution kernel weights, Represents the discrete sampling positions within the convolution kernel; function Perform a spatial translation operation to align the feature tensors with their respective grid coordinates; The feature representations of the two paths are respectively processed through learnable coefficients. and Weighted and scaled values ​​are then integrated via element-wise fusion before the P2 prediction layer to generate the final high-resolution output. 。

Images